Input understanding systems and services help users perform different tasks (whether executing an action or providing the correct information) by conducting a conversation with the user in which multiple turns are involved. However, the conversation does not necessarily traverse a fixed set of states. To fulfill any task, an input understanding system/service needs various pieces of information that may be elicited from the user. Collection of such information should be executed in a way that: 1) is efficient for the user; and 2) does not burden an executor of a task with task state tracking details. It is with respect to this general technical environment that the present application is directed.